kind of darkly funny that "gender studies" is the stereotypical "useless degree" because gender studies will help you understand a large and important chunk of the current psychosis in american life

No rush, I should be asleep anyway :) enjoy the gym, will check back tomorrow 🫡

It also immediately (unprompted) tried to write `NLP Research Project: Incorporating domain-specific knowledge into NLP models` in `/projects/nlp-research/project_description.txt`, and defaulted to directories in `/home/daniele`, which I guess are leftover hard-codings?

Ran into a few issues with `run.py`, which are addressed here: github.com/DanieleSalat...

I've decided to stop calling it "Twitter." Used to do that as an insult to Musk, a denial of his reality. But Twitter is dead and buried forever. It's just X, the Nazi site for Nazis, and that's all it will ever be. Get the fuck off.

I'm making no comment on whether or not that's feasible (it's probably not, but that's not the point). I'm asking why "mass-producing rockets" *as a theory* has anything to do with the Ideal Rocket Equation.

I am always entirely 1000% down for clowning on Elon, but I don't understand what you're saying or criticizing here. Isn't the quote about building many efficient-ish rockets to increase throughput, not about attempting to exceed a theoretical maximum efficiency?

Fascinating stuff, thank you for sharing! I'd been tinkering similarly over Christmas break, though you're lightyears ahead of me. Surprised & impressed that you got it working w/ Claude - I'd assumed I'd need a self-hosted model to control the system prompt sufficiently. Will install and tinker!

Oh shit, and they've already progressed beyond basic filesystem to vector dbs in github.com/mudler/Local.... Hell yeah.

LocalAI plus github.com/mudler/Local... looks promising

My main requirement is that the tool/agent/whatever runs locally (so has access to my local filesystem for "memory"). I'd _like_ for it to also be able to use a locally-served model, but that's not a requirement.

This seems like a generic-enough problem that many someones should have already solved it, or at least made more progress than me - "how do you inform a model of available tools, and how do you enforce that it actually uses them when it says it does?". Feels MCP-ish, but is that overkill?

Over the last couple weeks I've tinkered a bunch with the idea of "a chatbot with an intermediate layer that persists what it 'learns' to the filesystem, with a periodic job to reindex the knowledge base". Turns out many models will _say_ "I'll remember that" but not emit the tool call trigger

One of my intentions for the New Year is to get more curious and less judgemental about precisely this attitude. Would you mind if I asked you some questions in DM?

Lanie and I started watching the "The Fast And The Furious" series on Christmas Eve (a.k.a the most inconsistently-named series of all time). Only two movies in, it hasn't gotten to the truly absurd escalations yet, but I have hope.